granicus.if.org Git - postgresql/blob - src/backend/replication/logical/snapbuild.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * snapbuild.c
   4  *
   5  *        Infrastructure for building historic catalog snapshots based on contents
   6  *        of the WAL, for the purpose of decoding heapam.c style values in the
   7  *        WAL.
   8  *
   9  * NOTES:
  10  *
  11  * We build snapshots which can *only* be used to read catalog contents and we
  12  * do so by reading and interpreting the WAL stream. The aim is to build a
  13  * snapshot that behaves the same as a freshly taken MVCC snapshot would have
  14  * at the time the XLogRecord was generated.
  15  *
  16  * To build the snapshots we reuse the infrastructure built for Hot
  17  * Standby. The in-memory snapshots we build look different than HS' because
  18  * we have different needs. To successfully decode data from the WAL we only
  19  * need to access catalog tables and (sys|rel|cat)cache, not the actual user
  20  * tables since the data we decode is wholly contained in the WAL
  21  * records. Also, our snapshots need to be different in comparison to normal
  22  * MVCC ones because in contrast to those we cannot fully rely on the clog and
  23  * pg_subtrans for information about committed transactions because they might
  24  * commit in the future from the POV of the WAL entry we're currently
  25  * decoding. This definition has the advantage that we only need to prevent
  26  * removal of catalog rows, while normal table's rows can still be
  27  * removed. This is achieved by using the replication slot mechanism.
  28  *
  29  * As the percentage of transactions modifying the catalog normally is fairly
  30  * small in comparisons to ones only manipulating user data, we keep track of
  31  * the committed catalog modifying ones inside (xmin, xmax) instead of keeping
  32  * track of all running transactions like its done in a normal snapshot. Note
  33  * that we're generally only looking at transactions that have acquired an
  34  * xid. That is we keep a list of transactions between snapshot->(xmin, xmax)
  35  * that we consider committed, everything else is considered aborted/in
  36  * progress. That also allows us not to care about subtransactions before they
  37  * have committed which means this modules, in contrast to HS, doesn't have to
  38  * care about suboverflowed subtransactions and similar.
  39  *
  40  * One complexity of doing this is that to e.g. handle mixed DDL/DML
  41  * transactions we need Snapshots that see intermediate versions of the
  42  * catalog in a transaction. During normal operation this is achieved by using
  43  * CommandIds/cmin/cmax. The problem with that however is that for space
  44  * efficiency reasons only one value of that is stored
  45  * (c.f. combocid.c). Since ComboCids are only available in memory we log
  46  * additional information which allows us to get the original (cmin, cmax)
  47  * pair during visibility checks. Check the reorderbuffer.c's comment above
  48  * ResolveCminCmaxDuringDecoding() for details.
  49  *
  50  * To facilitate all this we need our own visibility routine, as the normal
  51  * ones are optimized for different usecases.
  52  *
  53  * To replace the normal catalog snapshots with decoding ones use the
  54  * SetupHistoricSnapshot() and TeardownHistoricSnapshot() functions.
  55  *
  56  *
  57  *
  58  * The snapbuild machinery is starting up in several stages, as illustrated
  59  * by the following graph:
  60  *         +-------------------------+
  61  *    +----|SNAPBUILD_START          |-------------+
  62  *    |    +-------------------------+             |
  63  *    |                 |                          |
  64  *    |                 |                          |
  65  *    |     running_xacts with running xacts       |
  66  *    |                 |                          |
  67  *    |                 |                          |
  68  *    |                 v                          |
  69  *    |    +-------------------------+             v
  70  *    |    |SNAPBUILD_FULL_SNAPSHOT  |------------>|
  71  *    |    +-------------------------+             |
  72  * running_xacts        |                      saved snapshot
  73  * with zero xacts      |                 at running_xacts's lsn
  74  *    |                 |                          |
  75  *    |     all running toplevel TXNs finished     |
  76  *    |                 |                          |
  77  *    |                 v                          |
  78  *    |    +-------------------------+             |
  79  *    +--->|SNAPBUILD_CONSISTENT     |<------------+
  80  *         +-------------------------+
  81  *
  82  * Initially the machinery is in the START stage. When a xl_running_xacts
  83  * record is read that is sufficiently new (above the safe xmin horizon),
  84  * there's a state transation. If there were no running xacts when the
  85  * runnign_xacts record was generated, we'll directly go into CONSISTENT
  86  * state, otherwise we'll switch to the FULL_SNAPSHOT state. Having a full
  87  * snapshot means that all transactions that start henceforth can be decoded
  88  * in their entirety, but transactions that started previously can't. In
  89  * FULL_SNAPSHOT we'll switch into CONSISTENT once all those previously
  90  * running transactions have committed or aborted.
  91  *
  92  * Only transactions that commit after CONSISTENT state has been reached will
  93  * be replayed, even though they might have started while still in
  94  * FULL_SNAPSHOT. That ensures that we'll reach a point where no previous
  95  * changes has been exported, but all the following ones will be. That point
  96  * is a convenient point to initialize replication from, which is why we
  97  * export a snapshot at that point, which *can* be used to read normal data.
  98  *
  99  * Copyright (c) 2012-2014, PostgreSQL Global Development Group
 100  *
 101  * IDENTIFICATION
 102  *        src/backend/replication/snapbuild.c
 103  *
 104  *-------------------------------------------------------------------------
 105  */
 106
 107 #include "postgres.h"
 108
 109 #include <sys/stat.h>
 110 #include <sys/types.h>
 111 #include <unistd.h>
 112
 113 #include "miscadmin.h"
 114
 115 #include "access/heapam_xlog.h"
 116 #include "access/transam.h"
 117 #include "access/xact.h"
 118
 119 #include "replication/logical.h"
 120 #include "replication/reorderbuffer.h"
 121 #include "replication/snapbuild.h"
 122
 123 #include "utils/builtins.h"
 124 #include "utils/memutils.h"
 125 #include "utils/snapshot.h"
 126 #include "utils/snapmgr.h"
 127 #include "utils/tqual.h"
 128
 129 #include "storage/block.h"              /* debugging output */
 130 #include "storage/fd.h"
 131 #include "storage/lmgr.h"
 132 #include "storage/proc.h"
 133 #include "storage/procarray.h"
 134 #include "storage/standby.h"
 135
 136 /*
 137  * This struct contains the current state of the snapshot building
 138  * machinery. Besides a forward declaration in the header, it is not exposed
 139  * to the public, so we can easily change it's contents.
 140  */
 141 struct SnapBuild
 142 {
 143         /* how far are we along building our first full snapshot */
 144         SnapBuildState state;
 145
 146         /* private memory context used to allocate memory for this module. */
 147         MemoryContext context;
 148
 149         /* all transactions < than this have committed/aborted */
 150         TransactionId xmin;
 151
 152         /* all transactions >= than this are uncommitted */
 153         TransactionId xmax;
 154
 155         /*
 156          * Don't replay commits from an LSN <= this LSN. This can be set
 157          * externally but it will also be advanced (never retreat) from within
 158          * snapbuild.c.
 159          */
 160         XLogRecPtr      transactions_after;
 161
 162         /*
 163          * Don't start decoding WAL until the "xl_running_xacts" information
 164          * indicates there are no running xids with a xid smaller than this.
 165          */
 166         TransactionId initial_xmin_horizon;
 167
 168         /*
 169          * Snapshot that's valid to see the catalog state seen at this moment.
 170          */
 171         Snapshot        snapshot;
 172
 173         /*
 174          * LSN of the last location we are sure a snapshot has been serialized to.
 175          */
 176         XLogRecPtr      last_serialized_snapshot;
 177
 178         /*
 179          * The reorderbuffer we need to update with usable snapshots et al.
 180          */
 181         ReorderBuffer *reorder;
 182
 183         /*
 184          * Information about initially running transactions
 185          *
 186          * When we start building a snapshot there already may be transactions in
 187          * progress.  Those are stored in running.xip.  We don't have enough
 188          * information about those to decode their contents, so until they are
 189          * finished (xcnt=0) we cannot switch to a CONSISTENT state.
 190          */
 191         struct
 192         {
 193                 /*
 194                  * As long as running.xcnt all XIDs < running.xmin and > running.xmax
 195                  * have to be checked whether they still are running.
 196                  */
 197                 TransactionId xmin;
 198                 TransactionId xmax;
 199
 200                 size_t          xcnt;           /* number of used xip entries */
 201                 size_t          xcnt_space; /* allocated size of xip */
 202                 TransactionId *xip;             /* running xacts array, xidComparator-sorted */
 203         }                       running;
 204
 205         /*
 206          * Array of transactions which could have catalog changes that committed
 207          * between xmin and xmax.
 208          */
 209         struct
 210         {
 211                 /* number of committed transactions */
 212                 size_t          xcnt;
 213
 214                 /* available space for committed transactions */
 215                 size_t          xcnt_space;
 216
 217                 /*
 218                  * Until we reach a CONSISTENT state, we record commits of all
 219                  * transactions, not just the catalog changing ones. Record when that
 220                  * changes so we know we cannot export a snapshot safely anymore.
 221                  */
 222                 bool            includes_all_transactions;
 223
 224                 /*
 225                  * Array of committed transactions that have modified the catalog.
 226                  *
 227                  * As this array is frequently modified we do *not* keep it in
 228                  * xidComparator order. Instead we sort the array when building &
 229                  * distributing a snapshot.
 230                  *
 231                  * TODO: It's unclear whether that reasoning has much merit. Every
 232                  * time we add something here after becoming consistent will also
 233                  * require distributing a snapshot. Storing them sorted would
 234                  * potentially also make it easier to purge (but more complicated wrt
 235                  * wraparound?). Should be improved if sorting while building the
 236                  * snapshot shows up in profiles.
 237                  */
 238                 TransactionId *xip;
 239         }                       committed;
 240 };
 241
 242 /*
 243  * Starting a transaction -- which we need to do while exporting a snapshot --
 244  * removes knowledge about the previously used resowner, so we save it here.
 245  */
 246 ResourceOwner SavedResourceOwnerDuringExport = NULL;
 247 bool ExportInProgress = false;
 248
 249 /* transaction state manipulation functions */
 250 static void SnapBuildEndTxn(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid);
 251
 252 /* ->running manipulation */
 253 static bool SnapBuildTxnIsRunning(SnapBuild *builder, TransactionId xid);
 254
 255 /* ->committed manipulation */
 256 static void SnapBuildPurgeCommittedTxn(SnapBuild *builder);
 257
 258 /* snapshot building/manipulation/distribution functions */
 259 static Snapshot SnapBuildBuildSnapshot(SnapBuild *builder, TransactionId xid);
 260
 261 static void SnapBuildFreeSnapshot(Snapshot snap);
 262
 263 static void SnapBuildSnapIncRefcount(Snapshot snap);
 264
 265 static void SnapBuildDistributeNewCatalogSnapshot(SnapBuild *builder, XLogRecPtr lsn);
 266
 267 /* xlog reading helper functions for SnapBuildProcessRecord */
 268 static bool SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running);
 269
 270 /* serialization functions */
 271 static void SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn);
 272 static bool SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn);
 273
 274
 275 /*
 276  * Allocate a new snapshot builder.
 277  *
 278  * xmin_horizon is the xid >=which we can be sure no catalog rows have been
 279  * removed, start_lsn is the LSN >= we want to replay commits.
 280  */
 281 SnapBuild *
 282 AllocateSnapshotBuilder(ReorderBuffer *reorder,
 283                                                 TransactionId xmin_horizon,
 284                                                 XLogRecPtr start_lsn)
 285 {
 286         MemoryContext context;
 287         MemoryContext oldcontext;
 288         SnapBuild  *builder;
 289
 290         /* allocate memory in own context, to have better accountability */
 291         context = AllocSetContextCreate(CurrentMemoryContext,
 292                                                                         "snapshot builder context",
 293                                                                         ALLOCSET_DEFAULT_MINSIZE,
 294                                                                         ALLOCSET_DEFAULT_INITSIZE,
 295                                                                         ALLOCSET_DEFAULT_MAXSIZE);
 296         oldcontext = MemoryContextSwitchTo(context);
 297
 298         builder = palloc0(sizeof(SnapBuild));
 299
 300         builder->state = SNAPBUILD_START;
 301         builder->context = context;
 302         builder->reorder = reorder;
 303         /* Other struct members initialized by zeroing via palloc0 above */
 304
 305         builder->committed.xcnt = 0;
 306         builder->committed.xcnt_space = 128;            /* arbitrary number */
 307         builder->committed.xip =
 308                 palloc0(builder->committed.xcnt_space * sizeof(TransactionId));
 309         builder->committed.includes_all_transactions = true;
 310         builder->committed.xip =
 311                 palloc0(builder->committed.xcnt_space * sizeof(TransactionId));
 312         builder->initial_xmin_horizon = xmin_horizon;
 313         builder->transactions_after = start_lsn;
 314
 315         MemoryContextSwitchTo(oldcontext);
 316
 317         return builder;
 318 }
 319
 320 /*
 321  * Free a snapshot builder.
 322  */
 323 void
 324 FreeSnapshotBuilder(SnapBuild *builder)
 325 {
 326         MemoryContext context = builder->context;
 327
 328         /* free snapshot explicitly, that contains some error checking */
 329         if (builder->snapshot != NULL)
 330         {
 331                 SnapBuildSnapDecRefcount(builder->snapshot);
 332                 builder->snapshot = NULL;
 333         }
 334
 335         /* other resources are deallocated via memory context reset */
 336         MemoryContextDelete(context);
 337 }
 338
 339 /*
 340  * Free an unreferenced snapshot that has previously been built by us.
 341  */
 342 static void
 343 SnapBuildFreeSnapshot(Snapshot snap)
 344 {
 345         /* make sure we don't get passed an external snapshot */
 346         Assert(snap->satisfies == HeapTupleSatisfiesHistoricMVCC);
 347
 348         /* make sure nobody modified our snapshot */
 349         Assert(snap->curcid == FirstCommandId);
 350         Assert(!snap->suboverflowed);
 351         Assert(!snap->takenDuringRecovery);
 352         Assert(snap->regd_count == 1);
 353
 354         /* slightly more likely, so it's checked even without c-asserts */
 355         if (snap->copied)
 356                 elog(ERROR, "cannot free a copied snapshot");
 357
 358         if (snap->active_count)
 359                 elog(ERROR, "cannot free an active snapshot");
 360
 361         pfree(snap);
 362 }
 363
 364 /*
 365  * In which state of snapshot building are we?
 366  */
 367 SnapBuildState
 368 SnapBuildCurrentState(SnapBuild *builder)
 369 {
 370         return builder->state;
 371 }
 372
 373 /*
 374  * Should the contents of transaction ending at 'ptr' be decoded?
 375  */
 376 bool
 377 SnapBuildXactNeedsSkip(SnapBuild *builder, XLogRecPtr ptr)
 378 {
 379         return ptr <= builder->transactions_after;
 380 }
 381
 382 /*
 383  * Increase refcount of a snapshot.
 384  *
 385  * This is used when handing out a snapshot to some external resource or when
 386  * adding a Snapshot as builder->snapshot.
 387  */
 388 static void
 389 SnapBuildSnapIncRefcount(Snapshot snap)
 390 {
 391         snap->active_count++;
 392 }
 393
 394 /*
 395  * Decrease refcount of a snapshot and free if the refcount reaches zero.
 396  *
 397  * Externally visible, so that external resources that have been handed an
 398  * IncRef'ed Snapshot can adjust its refcount easily.
 399  */
 400 void
 401 SnapBuildSnapDecRefcount(Snapshot snap)
 402 {
 403         /* make sure we don't get passed an external snapshot */
 404         Assert(snap->satisfies == HeapTupleSatisfiesHistoricMVCC);
 405
 406         /* make sure nobody modified our snapshot */
 407         Assert(snap->curcid == FirstCommandId);
 408         Assert(!snap->suboverflowed);
 409         Assert(!snap->takenDuringRecovery);
 410
 411         Assert(snap->regd_count == 1);
 412
 413         Assert(snap->active_count);
 414
 415         /* slightly more likely, so its checked even without casserts */
 416         if (snap->copied)
 417                 elog(ERROR, "cannot free a copied snapshot");
 418
 419         snap->active_count--;
 420         if (!snap->active_count)
 421                 SnapBuildFreeSnapshot(snap);
 422 }
 423
 424 /*
 425  * Build a new snapshot, based on currently committed catalog-modifying
 426  * transactions.
 427  *
 428  * In-progress transactions with catalog access are *not* allowed to modify
 429  * these snapshots; they have to copy them and fill in appropriate ->curcid
 430  * and ->subxip/subxcnt values.
 431  */
 432 static Snapshot
 433 SnapBuildBuildSnapshot(SnapBuild *builder, TransactionId xid)
 434 {
 435         Snapshot        snapshot;
 436         Size            ssize;
 437
 438         Assert(builder->state >= SNAPBUILD_FULL_SNAPSHOT);
 439
 440         ssize = sizeof(SnapshotData)
 441                 + sizeof(TransactionId) * builder->committed.xcnt
 442                 + sizeof(TransactionId) * 1 /* toplevel xid */ ;
 443
 444         snapshot = MemoryContextAllocZero(builder->context, ssize);
 445
 446         snapshot->satisfies = HeapTupleSatisfiesHistoricMVCC;
 447
 448         /*
 449          * We misuse the original meaning of SnapshotData's xip and subxip fields
 450          * to make the more fitting for our needs.
 451          *
 452          * In the 'xip' array we store transactions that have to be treated as
 453          * committed. Since we will only ever look at tuples from transactions
 454          * that have modified the catalog its more efficient to store those few
 455          * that exist between xmin and xmax (frequently there are none).
 456          *
 457          * Snapshots that are used in transactions that have modified the catalog
 458          * also use the 'subxip' array to store their toplevel xid and all the
 459          * subtransaction xids so we can recognize when we need to treat rows as
 460          * visible that are not in xip but still need to be visible. Subxip only
 461          * gets filled when the transaction is copied into the context of a
 462          * catalog modifying transaction since we otherwise share a snapshot
 463          * between transactions. As long as a txn hasn't modified the catalog it
 464          * doesn't need to treat any uncommitted rows as visible, so there is no
 465          * need for those xids.
 466          *
 467          * Both arrays are qsort'ed so that we can use bsearch() on them.
 468          */
 469         Assert(TransactionIdIsNormal(builder->xmin));
 470         Assert(TransactionIdIsNormal(builder->xmax));
 471
 472         snapshot->xmin = builder->xmin;
 473         snapshot->xmax = builder->xmax;
 474
 475         /* store all transactions to be treated as committed by this snapshot */
 476         snapshot->xip =
 477                 (TransactionId *) ((char *) snapshot + sizeof(SnapshotData));
 478         snapshot->xcnt = builder->committed.xcnt;
 479         memcpy(snapshot->xip,
 480                    builder->committed.xip,
 481                    builder->committed.xcnt * sizeof(TransactionId));
 482
 483         /* sort so we can bsearch() */
 484         qsort(snapshot->xip, snapshot->xcnt, sizeof(TransactionId), xidComparator);
 485
 486         /*
 487          * Initially, subxip is empty, i.e. it's a snapshot to be used by
 488          * transactions that don't modify the catalog. Will be filled by
 489          * ReorderBufferCopySnap() if necessary.
 490          */
 491         snapshot->subxcnt = 0;
 492         snapshot->subxip = NULL;
 493
 494         snapshot->suboverflowed = false;
 495         snapshot->takenDuringRecovery = false;
 496         snapshot->copied = false;
 497         snapshot->curcid = FirstCommandId;
 498         snapshot->active_count = 0;
 499         snapshot->regd_count = 1; /* mark as registered so nobody frees it */
 500
 501         return snapshot;
 502 }
 503
 504 /*
 505  * Export a snapshot so it can be set in another session with SET TRANSACTION
 506  * SNAPSHOT.
 507  *
 508  * For that we need to start a transaction in the current backend as the
 509  * importing side checks whether the source transaction is still open to make
 510  * sure the xmin horizon hasn't advanced since then.
 511  *
 512  * After that we convert a locally built snapshot into the normal variant
 513  * understood by HeapTupleSatisfiesMVCC et al.
 514  */
 515 const char *
 516 SnapBuildExportSnapshot(SnapBuild *builder)
 517 {
 518         Snapshot        snap;
 519         char       *snapname;
 520         TransactionId xid;
 521         TransactionId *newxip;
 522         int                     newxcnt = 0;
 523
 524         if (builder->state != SNAPBUILD_CONSISTENT)
 525                 elog(ERROR, "cannot export a snapshot before reaching a consistent state");
 526
 527         if (!builder->committed.includes_all_transactions)
 528                 elog(ERROR, "cannot export a snapshot, not all transactions are monitored anymore");
 529
 530         /* so we don't overwrite the existing value */
 531         if (TransactionIdIsValid(MyPgXact->xmin))
 532                 elog(ERROR, "cannot export a snapshot when MyPgXact->xmin already is valid");
 533
 534         if (IsTransactionOrTransactionBlock())
 535                 elog(ERROR, "cannot export a snapshot from within a transaction");
 536
 537         if (SavedResourceOwnerDuringExport)
 538                 elog(ERROR, "can only export one snapshot at a time");
 539
 540         SavedResourceOwnerDuringExport = CurrentResourceOwner;
 541         ExportInProgress = true;
 542
 543         StartTransactionCommand();
 544
 545         Assert(!FirstSnapshotSet);
 546
 547         /* There doesn't seem to a nice API to set these */
 548         XactIsoLevel = XACT_REPEATABLE_READ;
 549         XactReadOnly = true;
 550
 551         snap = SnapBuildBuildSnapshot(builder, GetTopTransactionId());
 552
 553         /*
 554          * We know that snap->xmin is alive, enforced by the logical xmin
 555          * mechanism. Due to that we can do this without locks, we're only
 556          * changing our own value.
 557          */
 558         MyPgXact->xmin = snap->xmin;
 559
 560         /* allocate in transaction context */
 561         newxip = (TransactionId *)
 562                 palloc(sizeof(TransactionId) * GetMaxSnapshotXidCount());
 563
 564         /*
 565          * snapbuild.c builds transactions in an "inverted" manner, which means it
 566          * stores committed transactions in ->xip, not ones in progress. Build a
 567          * classical snapshot by marking all non-committed transactions as
 568          * in-progress. This can be expensive.
 569          */
 570         for (xid = snap->xmin; NormalTransactionIdPrecedes(xid, snap->xmax);)
 571         {
 572                 void       *test;
 573
 574                 /*
 575                  * Check whether transaction committed using the decoding snapshot
 576                  * meaning of ->xip.
 577                  */
 578                 test = bsearch(&xid, snap->xip, snap->xcnt,
 579                                            sizeof(TransactionId), xidComparator);
 580
 581                 if (test == NULL)
 582                 {
 583                         if (newxcnt >= GetMaxSnapshotXidCount())
 584                                 elog(ERROR, "snapshot too large");
 585
 586                         newxip[newxcnt++] = xid;
 587                 }
 588
 589                 TransactionIdAdvance(xid);
 590         }
 591
 592         snap->xcnt = newxcnt;
 593         snap->xip = newxip;
 594
 595         /*
 596          * now that we've built a plain snapshot, use the normal mechanisms for
 597          * exporting it
 598          */
 599         snapname = ExportSnapshot(snap);
 600
 601         ereport(LOG,
 602                         (errmsg("exported logical decoding snapshot: \"%s\" with %u xids",
 603                                         snapname, snap->xcnt)));
 604         return snapname;
 605 }
 606
 607 /*
 608  * Reset a previously SnapBuildExportSnapshot()'ed snapshot if there is
 609  * any. Aborts the previously started transaction and resets the resource
 610  * owner back to it's original value.
 611  */
 612 void
 613 SnapBuildClearExportedSnapshot()
 614 {
 615         /* nothing exported, thats the usual case */
 616         if (!ExportInProgress)
 617                 return;
 618
 619         if (!IsTransactionState())
 620                 elog(ERROR, "clearing exported snapshot in wrong transaction state");
 621
 622         /* make sure nothing  could have ever happened */
 623         AbortCurrentTransaction();
 624
 625         CurrentResourceOwner = SavedResourceOwnerDuringExport;
 626         SavedResourceOwnerDuringExport = NULL;
 627         ExportInProgress = false;
 628 }
 629
 630 /*
 631  * Handle the effects of a single heap change, appropriate to the current state
 632  * of the snapshot builder and returns whether changes made at (xid, lsn) can
 633  * be decoded.
 634  */
 635 bool
 636 SnapBuildProcessChange(SnapBuild *builder, TransactionId xid, XLogRecPtr lsn)
 637 {
 638         bool is_old_tx;
 639
 640         /*
 641          * We can't handle data in transactions if we haven't built a snapshot
 642          * yet, so don't store them.
 643          */
 644         if (builder->state < SNAPBUILD_FULL_SNAPSHOT)
 645                 return false;
 646
 647         /*
 648          * No point in keeping track of changes in transactions that we don't have
 649          * enough information about to decode. This means that they started before
 650          * we got into the SNAPBUILD_FULL_SNAPSHOT state.
 651          */
 652         if (builder->state < SNAPBUILD_CONSISTENT &&
 653                 SnapBuildTxnIsRunning(builder, xid))
 654                 return false;
 655
 656         /*
 657          * If the reorderbuffer doesn't yet have a snapshot, add one now, it will
 658          * be needed to decode the change we're currently processing.
 659          */
 660         is_old_tx = ReorderBufferIsXidKnown(builder->reorder, xid);
 661
 662         if (!is_old_tx || !ReorderBufferXidHasBaseSnapshot(builder->reorder, xid))
 663         {
 664                 /* only build a new snapshot if we don't have a prebuilt one */
 665                 if (builder->snapshot == NULL)
 666                 {
 667                         builder->snapshot = SnapBuildBuildSnapshot(builder, xid);
 668                         /* inrease refcount for the snapshot builder */
 669                         SnapBuildSnapIncRefcount(builder->snapshot);
 670                 }
 671
 672                 /*
 673                  * Increase refcount for the transaction we're handing the snapshot
 674                  * out to.
 675                  */
 676                 SnapBuildSnapIncRefcount(builder->snapshot);
 677                 ReorderBufferSetBaseSnapshot(builder->reorder, xid, lsn,
 678                                                                          builder->snapshot);
 679         }
 680
 681         return true;
 682 }
 683
 684 /*
 685  * Do CommandId/ComboCid handling after reading a xl_heap_new_cid record. This
 686  * implies that a transaction has done some form of write to system catalogs.
 687  */
 688 void
 689 SnapBuildProcessNewCid(SnapBuild *builder, TransactionId xid,
 690                                            XLogRecPtr lsn, xl_heap_new_cid *xlrec)
 691 {
 692         CommandId       cid;
 693
 694         /*
 695          * we only log new_cid's if a catalog tuple was modified, so mark
 696          * the transaction as containing catalog modifications
 697          */
 698         ReorderBufferXidSetCatalogChanges(builder->reorder, xid,lsn);
 699
 700         ReorderBufferAddNewTupleCids(builder->reorder, xlrec->top_xid, lsn,
 701                                                                  xlrec->target.node, xlrec->target.tid,
 702                                                                  xlrec->cmin, xlrec->cmax,
 703                                                                  xlrec->combocid);
 704
 705         /* figure out new command id */
 706         if (xlrec->cmin != InvalidCommandId &&
 707                 xlrec->cmax != InvalidCommandId)
 708                 cid = Max(xlrec->cmin, xlrec->cmax);
 709         else if (xlrec->cmax != InvalidCommandId)
 710                 cid = xlrec->cmax;
 711         else if (xlrec->cmin != InvalidCommandId)
 712                 cid = xlrec->cmin;
 713         else
 714         {
 715                 cid = InvalidCommandId;         /* silence compiler */
 716                 elog(ERROR, "xl_heap_new_cid record without a valid CommandId");
 717         }
 718
 719         ReorderBufferAddNewCommandId(builder->reorder, xid, lsn, cid + 1);
 720 }
 721
 722 /*
 723  * Check whether `xid` is currently 'running'.
 724  *
 725  * Running transactions in our parlance are transactions which we didn't
 726  * observe from the start so we can't properly decode their contents. They
 727  * only exist after we freshly started from an < CONSISTENT snapshot.
 728  */
 729 static bool
 730 SnapBuildTxnIsRunning(SnapBuild *builder, TransactionId xid)
 731 {
 732         Assert(builder->state < SNAPBUILD_CONSISTENT);
 733         Assert(TransactionIdIsNormal(builder->running.xmin));
 734         Assert(TransactionIdIsNormal(builder->running.xmax));
 735
 736         if (builder->running.xcnt &&
 737                 NormalTransactionIdFollows(xid, builder->running.xmin) &&
 738                 NormalTransactionIdPrecedes(xid, builder->running.xmax))
 739         {
 740                 TransactionId *search =
 741                 bsearch(&xid, builder->running.xip, builder->running.xcnt_space,
 742                                 sizeof(TransactionId), xidComparator);
 743
 744                 if (search != NULL)
 745                 {
 746                         Assert(*search == xid);
 747                         return true;
 748                 }
 749         }
 750
 751         return false;
 752 }
 753
 754 /*
 755  * Add a new Snapshot to all transactions we're decoding that currently are
 756  * in-progress so they can see new catalog contents made by the transaction
 757  * that just committed. This is necessary because those in-progress
 758  * transactions will use the new catalog's contents from here on (at the very
 759  * least everything they do needs to be compatible with newer catalog
 760  * contents).
 761  */
 762 static void
 763 SnapBuildDistributeNewCatalogSnapshot(SnapBuild *builder, XLogRecPtr lsn)
 764 {
 765         dlist_iter      txn_i;
 766         ReorderBufferTXN *txn;
 767
 768         /*
 769          * Iterate through all toplevel transactions. This can include
 770          * subtransactions which we just don't yet know to be that, but that's
 771          * fine, they will just get an unneccesary snapshot queued.
 772          */
 773         dlist_foreach(txn_i, &builder->reorder->toplevel_by_lsn)
 774         {
 775                 txn = dlist_container(ReorderBufferTXN, node, txn_i.cur);
 776
 777                 Assert(TransactionIdIsValid(txn->xid));
 778
 779                 /*
 780                  * If we don't have a base snapshot yet, there are no changes in this
 781                  * transaction which in turn implies we don't yet need a snapshot at
 782                  * all. We'll add a snapshot when the first change gets queued.
 783                  *
 784                  * NB: This works correctly even for subtransactions because
 785                  * ReorderBufferCommitChild() takes care to pass the parent the base
 786                  * snapshot, and while iterating the changequeue we'll get the change
 787                  * from the subtxn.
 788                  */
 789                 if (!ReorderBufferXidHasBaseSnapshot(builder->reorder, txn->xid))
 790                         continue;
 791
 792                 elog(DEBUG2, "adding a new snapshot to %u at %X/%X",
 793                          txn->xid, (uint32) (lsn >> 32), (uint32) lsn);
 794
 795                 /*
 796                  * increase the snapshot's refcount for the transaction we are handing
 797                  * it out to
 798                  */
 799                 SnapBuildSnapIncRefcount(builder->snapshot);
 800                 ReorderBufferAddSnapshot(builder->reorder, txn->xid, lsn,
 801                                                                  builder->snapshot);
 802         }
 803 }
 804
 805 /*
 806  * Keep track of a new catalog changing transaction that has committed.
 807  */
 808 static void
 809 SnapBuildAddCommittedTxn(SnapBuild *builder, TransactionId xid)
 810 {
 811         Assert(TransactionIdIsValid(xid));
 812
 813         if (builder->committed.xcnt == builder->committed.xcnt_space)
 814         {
 815                 builder->committed.xcnt_space = builder->committed.xcnt_space * 2 + 1;
 816
 817                 elog(DEBUG1, "increasing space for committed transactions to %u",
 818                          (uint32) builder->committed.xcnt_space);
 819
 820                 builder->committed.xip = repalloc(builder->committed.xip,
 821                                         builder->committed.xcnt_space * sizeof(TransactionId));
 822         }
 823
 824         /*
 825          * TODO: It might make sense to keep the array sorted here instead of
 826          * doing it every time we build a new snapshot. On the other hand this
 827          * gets called repeatedly when a transaction with subtransactions commits.
 828          */
 829         builder->committed.xip[builder->committed.xcnt++] = xid;
 830 }
 831
 832 /*
 833  * Remove knowledge about transactions we treat as committed that are smaller
 834  * than ->xmin. Those won't ever get checked via the ->commited array but via
 835  * the clog machinery, so we don't need to waste memory on them.
 836  */
 837 static void
 838 SnapBuildPurgeCommittedTxn(SnapBuild *builder)
 839 {
 840         int                     off;
 841         TransactionId *workspace;
 842         int                     surviving_xids = 0;
 843
 844         /* not ready yet */
 845         if (!TransactionIdIsNormal(builder->xmin))
 846                 return;
 847
 848         /* TODO: Neater algorithm than just copying and iterating? */
 849         workspace =
 850                 MemoryContextAlloc(builder->context,
 851                                                    builder->committed.xcnt * sizeof(TransactionId));
 852
 853         /* copy xids that still are interesting to workspace */
 854         for (off = 0; off < builder->committed.xcnt; off++)
 855         {
 856                 if (NormalTransactionIdPrecedes(builder->committed.xip[off],
 857                                                                                 builder->xmin))
 858                         ;                                       /* remove */
 859                 else
 860                         workspace[surviving_xids++] = builder->committed.xip[off];
 861         }
 862
 863         /* copy workspace back to persistent state */
 864         memcpy(builder->committed.xip, workspace,
 865                    surviving_xids * sizeof(TransactionId));
 866
 867         elog(DEBUG3, "purged committed transactions from %u to %u, xmin: %u, xmax: %u",
 868                  (uint32) builder->committed.xcnt, (uint32) surviving_xids,
 869                  builder->xmin, builder->xmax);
 870         builder->committed.xcnt = surviving_xids;
 871
 872         pfree(workspace);
 873 }
 874
 875 /*
 876  * Common logic for SnapBuildAbortTxn and SnapBuildCommitTxn dealing with
 877  * keeping track of the amount of running transactions.
 878  */
 879 static void
 880 SnapBuildEndTxn(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid)
 881 {
 882         if (builder->state == SNAPBUILD_CONSISTENT)
 883                 return;
 884
 885         /*
 886          * NB: This handles subtransactions correctly even if we started from
 887          * suboverflowed xl_running_xacts because we only keep track of toplevel
 888          * transactions. Since the latter are always are allocated before their
 889          * subxids and since they end at the same time it's sufficient to deal
 890          * with them here.
 891          */
 892         if (SnapBuildTxnIsRunning(builder, xid))
 893         {
 894                 Assert(builder->running.xcnt > 0);
 895
 896                 if (!--builder->running.xcnt)
 897                 {
 898                         /*
 899                          * None of the originally running transaction is running anymore,
 900                          * so our incrementaly built snapshot now is consistent.
 901                          */
 902                         ereport(LOG,
 903                                         (errmsg("logical decoding found consistent point at %X/%X",
 904                                                         (uint32)(lsn >> 32), (uint32)lsn),
 905                                          errdetail("xid %u finished, no running transactions anymore",
 906                                                            xid)));
 907                         builder->state = SNAPBUILD_CONSISTENT;
 908                 }
 909         }
 910 }
 911
 912 /*
 913  * Abort a transaction, throw away all state we kept.
 914  */
 915 void
 916 SnapBuildAbortTxn(SnapBuild *builder, XLogRecPtr lsn,
 917                                   TransactionId xid,
 918                                   int nsubxacts, TransactionId *subxacts)
 919 {
 920         int                     i;
 921
 922         for (i = 0; i < nsubxacts; i++)
 923         {
 924                 TransactionId subxid = subxacts[i];
 925
 926                 SnapBuildEndTxn(builder, lsn, subxid);
 927         }
 928
 929         SnapBuildEndTxn(builder, lsn, xid);
 930 }
 931
 932 /*
 933  * Handle everything that needs to be done when a transaction commits
 934  */
 935 void
 936 SnapBuildCommitTxn(SnapBuild *builder, XLogRecPtr lsn, TransactionId xid,
 937                                    int nsubxacts, TransactionId *subxacts)
 938 {
 939         int                     nxact;
 940
 941         bool            forced_timetravel = false;
 942         bool            sub_needs_timetravel = false;
 943         bool            top_needs_timetravel = false;
 944
 945         TransactionId xmax = xid;
 946
 947         /*
 948          * If we couldn't observe every change of a transaction because it was
 949          * already running at the point we started to observe we have to assume it
 950          * made catalog changes.
 951          *
 952          * This has the positive benefit that we afterwards have enough
 953          * information to build an exportable snapshot that's usable by pg_dump et
 954          * al.
 955          */
 956         if (builder->state < SNAPBUILD_CONSISTENT)
 957         {
 958                 /* ensure that only commits after this are getting replayed */
 959                 if (builder->transactions_after < lsn)
 960                         builder->transactions_after = lsn;
 961
 962                 /*
 963                  * We could avoid treating !SnapBuildTxnIsRunning transactions as
 964                  * timetravel ones, but we want to be able to export a snapshot when
 965                  * we reached consistency.
 966                  */
 967                 forced_timetravel = true;
 968                 elog(DEBUG1, "forced to assume catalog changes for xid %u because it was running to early", xid);
 969         }
 970
 971         for (nxact = 0; nxact < nsubxacts; nxact++)
 972         {
 973                 TransactionId subxid = subxacts[nxact];
 974
 975                 /*
 976                  * make sure txn is not tracked in running txn's anymore, switch state
 977                  */
 978                 SnapBuildEndTxn(builder, lsn, subxid);
 979
 980                 /*
 981                  * If we're forcing timetravel we also need visibility information
 982                  * about subtransaction, so keep track of subtransaction's state.
 983                  */
 984                 if (forced_timetravel)
 985                 {
 986                         SnapBuildAddCommittedTxn(builder, subxid);
 987                         if (NormalTransactionIdFollows(subxid, xmax))
 988                                 xmax = subxid;
 989                 }
 990
 991                 /*
 992                  * Add subtransaction to base snapshot if it DDL, we don't distinguish
 993                  * to toplevel transactions there.
 994                  */
 995                 else if (ReorderBufferXidHasCatalogChanges(builder->reorder, subxid))
 996                 {
 997                         sub_needs_timetravel = true;
 998
 999                         elog(DEBUG1, "found subtransaction %u:%u with catalog changes.",
1000                                  xid, subxid);
1001
1002                         SnapBuildAddCommittedTxn(builder, subxid);
1003
1004                         if (NormalTransactionIdFollows(subxid, xmax))
1005                                 xmax = subxid;
1006                 }
1007         }
1008
1009         /*
1010          * Make sure toplevel txn is not tracked in running txn's anymore, switch
1011          * state to consistent if possible.
1012          */
1013         SnapBuildEndTxn(builder, lsn, xid);
1014
1015         if (forced_timetravel)
1016         {
1017                 elog(DEBUG2, "forced transaction %u to do timetravel.", xid);
1018
1019                 SnapBuildAddCommittedTxn(builder, xid);
1020         }
1021         /* add toplevel transaction to base snapshot */
1022         else if (ReorderBufferXidHasCatalogChanges(builder->reorder, xid))
1023         {
1024                 elog(DEBUG2, "found top level transaction %u, with catalog changes!",
1025                          xid);
1026
1027                 top_needs_timetravel = true;
1028                 SnapBuildAddCommittedTxn(builder, xid);
1029         }
1030         else if (sub_needs_timetravel)
1031         {
1032                 /* mark toplevel txn as timetravel as well */
1033                 SnapBuildAddCommittedTxn(builder, xid);
1034         }
1035
1036         /* if there's any reason to build a historic snapshot, to so now */
1037         if (forced_timetravel || top_needs_timetravel || sub_needs_timetravel)
1038         {
1039                 /*
1040                  * Adjust xmax of the snapshot builder, we only do that for committed,
1041                  * catalog modifying, transactions, everything else isn't interesting
1042                  * for us since we'll never look at the respective rows.
1043                  */
1044                 if (!TransactionIdIsValid(builder->xmax) ||
1045                         TransactionIdFollowsOrEquals(xmax, builder->xmax))
1046                 {
1047                         builder->xmax = xmax;
1048                         TransactionIdAdvance(builder->xmax);
1049                 }
1050
1051                 /*
1052                  * If we haven't built a complete snapshot yet there's no need to hand
1053                  * it out, it wouldn't (and couldn't) be used anyway.
1054                  */
1055                 if (builder->state < SNAPBUILD_FULL_SNAPSHOT)
1056                         return;
1057
1058                 /*
1059                  * Decrease the snapshot builder's refcount of the old snapshot, note
1060                  * that it still will be used if it has been handed out to the
1061                  * reorderbuffer earlier.
1062                  */
1063                 if (builder->snapshot)
1064                         SnapBuildSnapDecRefcount(builder->snapshot);
1065
1066                 builder->snapshot = SnapBuildBuildSnapshot(builder, xid);
1067
1068                 /* we might need to execute invalidations, add snapshot */
1069                 if (!ReorderBufferXidHasBaseSnapshot(builder->reorder, xid))
1070                 {
1071                         SnapBuildSnapIncRefcount(builder->snapshot);
1072                         ReorderBufferSetBaseSnapshot(builder->reorder, xid, lsn,
1073                                                                                  builder->snapshot);
1074                 }
1075
1076                 /* refcount of the snapshot builder for the new snapshot */
1077                 SnapBuildSnapIncRefcount(builder->snapshot);
1078
1079                 /* add a new SnapshotNow to all currently running transactions */
1080                 SnapBuildDistributeNewCatalogSnapshot(builder, lsn);
1081         }
1082         else
1083         {
1084                 /* record that we cannot export a general snapshot anymore */
1085                 builder->committed.includes_all_transactions = false;
1086         }
1087 }
1088
1089
1090 /* -----------------------------------
1091  * Snapshot building functions dealing with xlog records
1092  * -----------------------------------
1093  */
1094
1095 /*
1096  * Process a running xacts record, and use it's information to first build a
1097  * historic snapshot and later to release resources that aren't needed
1098  * anymore.
1099  */
1100 void
1101 SnapBuildProcessRunningXacts(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running)
1102 {
1103         ReorderBufferTXN *txn;
1104
1105         /*
1106          * If we're not consistent yet, inspect the record to see whether it
1107          * allows to get closer to being consistent. If we are consistent, dump
1108          * our snapshot so others or we, after a restart, can use it.
1109          */
1110         if (builder->state < SNAPBUILD_CONSISTENT)
1111         {
1112                 /* returns false if there's no point in performing cleanup just yet */
1113                 if (!SnapBuildFindSnapshot(builder, lsn, running))
1114                         return;
1115         }
1116         else
1117                 SnapBuildSerialize(builder, lsn);
1118
1119         /*
1120          * Update range of interesting xids base don the running xacts
1121          * information. We don't increase ->xmax using it, because once we are in
1122          * a consistent state we can do that ourselves and much more efficiently
1123          * so, because we only need to do it for catalog transactions since we
1124          * only ever look at those.
1125          *
1126          * NB: Because of that xmax can be lower than xmin, because we only
1127          * increase xmax when a catalog modifying transaction commits. While odd
1128          * looking, its correct and actually more efficient this way since we hit
1129          * fast paths in tqual.c.
1130          */
1131         builder->xmin = running->oldestRunningXid;
1132
1133         /* Remove transactions we don't need to keep track off anymore */
1134         SnapBuildPurgeCommittedTxn(builder);
1135
1136         elog(DEBUG3, "xmin: %u, xmax: %u, oldestrunning: %u",
1137                  builder->xmin, builder->xmax,
1138                  running->oldestRunningXid);
1139
1140         /*
1141          * Inrease shared memory limits, so vacuum can work on tuples we prevented
1142          * from being pruned till now.
1143          */
1144         LogicalIncreaseXminForSlot(lsn, running->oldestRunningXid);
1145
1146         /*
1147          * Also tell the slot where we can restart decoding from. We don't want to
1148          * do that after every commit because changing that implies an fsync of
1149          * the logical slot's state file, so we only do it every time we see a
1150          * running xacts record.
1151          *
1152          * Do so by looking for the oldest in progress transaction (determined by
1153          * the first LSN of any of its relevant records). Every transaction
1154          * remembers the last location we stored the snapshot to disk before its
1155          * beginning. That point is where we can restart from.
1156          */
1157
1158         /*
1159          * Can't know about a serialized snapshot's location if we're not
1160          * consistent.
1161          */
1162         if (builder->state < SNAPBUILD_CONSISTENT)
1163                 return;
1164
1165         txn = ReorderBufferGetOldestTXN(builder->reorder);
1166
1167         /*
1168          * oldest ongoing txn might have started when we didn't yet serialize
1169          * anything because we hadn't reached a consistent state yet.
1170          */
1171         if (txn != NULL && txn->restart_decoding_lsn != InvalidXLogRecPtr)
1172                 LogicalIncreaseRestartDecodingForSlot(lsn, txn->restart_decoding_lsn);
1173         /*
1174          * No in-progress transaction, can reuse the last serialized snapshot if
1175          * we have one.
1176          */
1177         else if (txn == NULL &&
1178                          builder->reorder->current_restart_decoding_lsn != InvalidXLogRecPtr &&
1179                          builder->last_serialized_snapshot != InvalidXLogRecPtr)
1180                 LogicalIncreaseRestartDecodingForSlot(lsn,
1181                                                                                    builder->last_serialized_snapshot);
1182 }
1183
1184
1185 /*
1186  * Build the start of a snapshot that's capable of decoding the catalog.
1187  *
1188  * Helper function for SnapBuildProcessRunningXacts() while we're not yet
1189  * consistent.
1190  *
1191  * Returns true if there is a point in performing internal maintenance/cleanup
1192  * using the xl_running_xacts record.
1193  */
1194 static bool
1195 SnapBuildFindSnapshot(SnapBuild *builder, XLogRecPtr lsn, xl_running_xacts *running)
1196 {
1197         /* ---
1198          * Build catalog decoding snapshot incrementally using information about
1199          * the currently running transactions. There are several ways to do that:
1200          *
1201          * a) There were no running transactions when the xl_running_xacts record
1202          *    was inserted, jump to CONSISTENT immediately. We might find such a
1203          *    state we were waiting for b) and c).
1204          *
1205          * b) Wait for all toplevel transactions that were running to end. We
1206          *    simply track the number of in-progress toplevel transactions and
1207          *    lower it whenever one commits or aborts. When that number
1208          *    (builder->running.xcnt) reaches zero, we can go from FULL_SNAPSHOT
1209          *    to CONSISTENT.
1210          *        NB: We need to search running.xip when seeing a transaction's end to
1211          *    make sure it's a toplevel transaction and it's been one of the
1212          *    intially running ones.
1213          *        Interestingly, in contrast to HS, this allows us not to care about
1214          *        subtransactions - and by extension suboverflowed xl_running_xacts -
1215          *        at all.
1216          *
1217          * c) This (in a previous run) or another decoding slot serialized a
1218          *    snapshot to disk that we can use.
1219          * ---
1220          */
1221
1222         /*
1223          * xl_running_xact record is older than what we can use, we might not have
1224          * all necessary catalog rows anymore.
1225          */
1226         if (TransactionIdIsNormal(builder->initial_xmin_horizon) &&
1227                 NormalTransactionIdPrecedes(running->oldestRunningXid,
1228                                                                         builder->initial_xmin_horizon))
1229         {
1230                 ereport(DEBUG1,
1231                                 (errmsg("skipping snapshot at %X/%X while building logical decoding snapshot, xmin horizon too low",
1232                                                 (uint32) (lsn >> 32), (uint32) lsn),
1233                                  errdetail("initial xmin horizon of %u vs the snapshot's %u",
1234                                                    builder->initial_xmin_horizon, running->oldestRunningXid)));
1235                 return true;
1236         }
1237
1238         /*
1239          * a) No transaction were running, we can jump to consistent.
1240          *
1241          * NB: We might have already started to incrementally assemble a snapshot,
1242          * so we need to be careful to deal with that.
1243          */
1244         if (running->xcnt == 0)
1245         {
1246                 if (builder->transactions_after == InvalidXLogRecPtr ||
1247                         builder->transactions_after < lsn)
1248                         builder->transactions_after = lsn;
1249
1250                 builder->xmin = running->oldestRunningXid;
1251                 builder->xmax = running->latestCompletedXid;
1252                 TransactionIdAdvance(builder->xmax);
1253
1254                 Assert(TransactionIdIsNormal(builder->xmin));
1255                 Assert(TransactionIdIsNormal(builder->xmax));
1256
1257                 /* no transactions running now */
1258                 builder->running.xcnt = 0;
1259                 builder->running.xmin = InvalidTransactionId;
1260                 builder->running.xmax = InvalidTransactionId;
1261
1262                 builder->state = SNAPBUILD_CONSISTENT;
1263
1264                 ereport(LOG,
1265                                 (errmsg("logical decoding found consistent point at %X/%X",
1266                                                 (uint32)(lsn >> 32), (uint32)lsn),
1267                                  errdetail("running xacts with xcnt == 0")));
1268
1269                 return false;
1270         }
1271         /* c) valid on disk state */
1272         else if (SnapBuildRestore(builder, lsn))
1273         {
1274                 /* there won't be any state to cleanup */
1275                 return false;
1276         }
1277         /*
1278          * b) first encounter of a useable xl_running_xacts record. If we had
1279          * found one earlier we would either track running transactions
1280          * (i.e. builder->running.xcnt != 0) or be consistent (this function
1281          * wouldn't get called).
1282          */
1283         else if (!builder->running.xcnt)
1284         {
1285                 int off;
1286
1287                 /*
1288                  * We only care about toplevel xids as those are the ones we
1289                  * definitely see in the wal stream. As snapbuild.c tracks committed
1290                  * instead of running transactions we don't need to know anything
1291                  * about uncommitted subtransactions.
1292                  */
1293                 builder->xmin = running->oldestRunningXid;
1294                 builder->xmax = running->latestCompletedXid;
1295                 TransactionIdAdvance(builder->xmax);
1296
1297                 /* so we can safely use the faster comparisons */
1298                 Assert(TransactionIdIsNormal(builder->xmin));
1299                 Assert(TransactionIdIsNormal(builder->xmax));
1300
1301                 builder->running.xcnt = running->xcnt;
1302                 builder->running.xcnt_space = running->xcnt;
1303                 builder->running.xip =
1304                         MemoryContextAlloc(builder->context,
1305                                                         builder->running.xcnt * sizeof(TransactionId));
1306                 memcpy(builder->running.xip, running->xids,
1307                            builder->running.xcnt * sizeof(TransactionId));
1308
1309                 /* sort so we can do a binary search */
1310                 qsort(builder->running.xip, builder->running.xcnt,
1311                           sizeof(TransactionId), xidComparator);
1312
1313                 builder->running.xmin = builder->running.xip[0];
1314                 builder->running.xmax = builder->running.xip[running->xcnt - 1];
1315
1316                 /* makes comparisons cheaper later */
1317                 TransactionIdRetreat(builder->running.xmin);
1318                 TransactionIdAdvance(builder->running.xmax);
1319
1320                 builder->state = SNAPBUILD_FULL_SNAPSHOT;
1321
1322                 ereport(LOG,
1323                                 (errmsg("logical decoding found initial starting point at %X/%X",
1324                                                 (uint32)(lsn >> 32), (uint32)lsn),
1325                                  errdetail("%u xacts need to finish", (uint32) builder->running.xcnt)));
1326
1327                 /*
1328                  * Iterate through all xids, wait for them to finish.
1329                  *
1330                  * This isn't required for the correctness of decoding, but to allow
1331                  * isolationtester to notice that we're currently waiting for
1332                  * something.
1333                  */
1334                 for(off = 0; off < builder->running.xcnt; off++)
1335                 {
1336                         TransactionId xid = builder->running.xip[off];
1337
1338                         /*
1339                          * Upper layers should prevent that we ever need to wait on
1340                          * ourselves. Check anyway, since failing to do so would either
1341                          * result in an endless wait or an Assert() failure.
1342                          */
1343                         if (TransactionIdIsCurrentTransactionId(xid))
1344                                 elog(ERROR, "waiting for ourselves");
1345
1346                         XactLockTableWait(xid);
1347                 }
1348
1349                 /* nothing could have built up so far, so don't perform cleanup */
1350                 return false;
1351         }
1352
1353         /*
1354          * We already started to track running xacts and need to wait for all
1355          * in-progress ones to finish. We fall through to the normal processing of
1356          * records so incremental cleanup can be performed.
1357          */
1358         return true;
1359 }
1360
1361
1362 /* -----------------------------------
1363  * Snapshot serialization support
1364  * -----------------------------------
1365  */
1366
1367 /*
1368  * We store current state of struct SnapBuild on disk in the following manner:
1369  *
1370  * struct SnapBuildOnDisk;
1371  * TransactionId * running.xcnt_space;
1372  * TransactionId * committed.xcnt; (*not xcnt_space*)
1373  *
1374  */
1375 typedef struct SnapBuildOnDisk
1376 {
1377         /* first part of this struct needs to be version independent */
1378
1379         /* data not covered by checksum */
1380         uint32          magic;
1381         pg_crc32        checksum;
1382
1383         /* data covered by checksum */
1384
1385         /* version, in case we want to support pg_upgrade */
1386         uint32          version;
1387         /* how large is the on disk data, excluding the constant sized part */
1388         uint32          length;
1389
1390         /* version dependent part */
1391         SnapBuild       builder;
1392
1393         /* variable amount of TransactionIds follows */
1394 } SnapBuildOnDisk;
1395
1396 #define SnapBuildOnDiskConstantSize \
1397         offsetof(SnapBuildOnDisk, builder)
1398 #define SnapBuildOnDiskNotChecksummedSize \
1399         offsetof(SnapBuildOnDisk, version)
1400
1401 #define SNAPBUILD_MAGIC 0x51A1E001
1402 #define SNAPBUILD_VERSION 1
1403
1404 /*
1405  * Store/Load a snapshot from disk, depending on the snapshot builder's state.
1406  *
1407  * Supposed to be used by external (i.e. not snapbuild.c) code that just reada
1408  * record that's a potential location for a serialized snapshot.
1409  */
1410 void
1411 SnapBuildSerializationPoint(SnapBuild *builder, XLogRecPtr lsn)
1412 {
1413         if (builder->state < SNAPBUILD_CONSISTENT)
1414                 SnapBuildRestore(builder, lsn);
1415         else
1416                 SnapBuildSerialize(builder, lsn);
1417 }
1418
1419 /*
1420  * Serialize the snapshot 'builder' at the location 'lsn' if it hasn't already
1421  * been done by another decoding process.
1422  */
1423 static void
1424 SnapBuildSerialize(SnapBuild *builder, XLogRecPtr lsn)
1425 {
1426         Size            needed_length;
1427         SnapBuildOnDisk *ondisk;
1428         char       *ondisk_c;
1429         int                     fd;
1430         char            tmppath[MAXPGPATH];
1431         char            path[MAXPGPATH];
1432         int                     ret;
1433         struct stat stat_buf;
1434         uint32          sz;
1435
1436         Assert(lsn != InvalidXLogRecPtr);
1437         Assert(builder->last_serialized_snapshot == InvalidXLogRecPtr ||
1438                    builder->last_serialized_snapshot <= lsn);
1439
1440         /*
1441          * no point in serializing if we cannot continue to work immediately after
1442          * restoring the snapshot
1443          */
1444         if (builder->state < SNAPBUILD_CONSISTENT)
1445                 return;
1446
1447         /*
1448          * We identify snapshots by the LSN they are valid for. We don't need to
1449          * include timelines in the name as each LSN maps to exactly one timeline
1450          * unless the user used pg_resetxlog or similar. If a user did so, there's
1451          * no hope continuing to decode anyway.
1452          */
1453         sprintf(path, "pg_llog/snapshots/%X-%X.snap",
1454                         (uint32) (lsn >> 32), (uint32) lsn);
1455
1456         /*
1457          * first check whether some other backend already has written the snapshot
1458          * for this LSN. It's perfectly fine if there's none, so we accept ENOENT
1459          * as a valid state. Everything else is an unexpected error.
1460          */
1461         ret = stat(path, &stat_buf);
1462
1463         if (ret != 0 && errno != ENOENT)
1464                 ereport(ERROR,
1465                                 (errmsg("could not stat file \"%s\": %m", path)));
1466
1467         else if (ret == 0)
1468         {
1469                 /*
1470                  * somebody else has already serialized to this point, don't overwrite
1471                  * but remember location, so we don't need to read old data again.
1472                  *
1473                  * To be sure it has been synced to disk after the rename() from the
1474                  * tempfile filename to the real filename, we just repeat the
1475                  * fsync. That ought to be cheap because in most scenarios it should
1476                  * already be safely on disk.
1477                  */
1478                 fsync_fname(path, false);
1479                 fsync_fname("pg_llog/snapshots", true);
1480
1481                 builder->last_serialized_snapshot = lsn;
1482                 goto out;
1483         }
1484
1485         /*
1486          * there is an obvious race condition here between the time we stat(2) the
1487          * file and us writing the file. But we rename the file into place
1488          * atomically and all files created need to contain the same data anyway,
1489          * so this is perfectly fine, although a bit of a resource waste. Locking
1490          * seems like pointless complication.
1491          */
1492         elog(DEBUG1, "serializing snapshot to %s", path);
1493
1494         /* to make sure only we will write to this tempfile, include pid */
1495         sprintf(tmppath, "pg_llog/snapshots/%X-%X.snap.%u.tmp",
1496                         (uint32) (lsn >> 32), (uint32) lsn, MyProcPid);
1497
1498         /*
1499          * Unlink temporary file if it already exists, needs to have been before a
1500          * crash/error since we won't enter this function twice from within a
1501          * single decoding slot/backend and the temporary file contains the pid of
1502          * the current process.
1503          */
1504         if (unlink(tmppath) != 0 && errno != ENOENT)
1505                 ereport(ERROR,
1506                                 (errcode_for_file_access(),
1507                                  errmsg("could not unlink file \"%s\": %m",     path)));
1508
1509         needed_length = sizeof(SnapBuildOnDisk) +
1510                 sizeof(TransactionId) * builder->running.xcnt_space +
1511                 sizeof(TransactionId) * builder->committed.xcnt;
1512
1513         ondisk_c = MemoryContextAllocZero(builder->context, needed_length);
1514         ondisk = (SnapBuildOnDisk *) ondisk_c;
1515         ondisk->magic = SNAPBUILD_MAGIC;
1516         ondisk->version = SNAPBUILD_VERSION;
1517         ondisk->length = needed_length;
1518         INIT_CRC32(ondisk->checksum);
1519         COMP_CRC32(ondisk->checksum,
1520                            ((char *) ondisk) + SnapBuildOnDiskNotChecksummedSize,
1521                            SnapBuildOnDiskConstantSize - SnapBuildOnDiskNotChecksummedSize);
1522         ondisk_c += sizeof(SnapBuildOnDisk);
1523
1524         memcpy(&ondisk->builder, builder, sizeof(SnapBuild));
1525         /* NULL-ify memory-only data */
1526         ondisk->builder.context = NULL;
1527         ondisk->builder.snapshot = NULL;
1528         ondisk->builder.reorder = NULL;
1529         ondisk->builder.running.xip = NULL;
1530         ondisk->builder.committed.xip = NULL;
1531
1532         COMP_CRC32(ondisk->checksum,
1533                            &ondisk->builder,
1534                            sizeof(SnapBuild));
1535
1536         /* copy running xacts */
1537         sz = sizeof(TransactionId) * builder->running.xcnt_space;
1538         memcpy(ondisk_c, builder->running.xip, sz);
1539         COMP_CRC32(ondisk->checksum, ondisk_c, sz);
1540         ondisk_c += sz;
1541
1542         /* copy committed xacts */
1543         sz = sizeof(TransactionId) * builder->committed.xcnt;
1544         memcpy(ondisk_c, builder->committed.xip, sz);
1545         COMP_CRC32(ondisk->checksum, ondisk_c, sz);
1546         ondisk_c += sz;
1547
1548         /* we have valid data now, open tempfile and write it there */
1549         fd = OpenTransientFile(tmppath,
1550                                                    O_CREAT | O_EXCL | O_WRONLY | PG_BINARY,
1551                                                    S_IRUSR | S_IWUSR);
1552         if (fd < 0)
1553                 ereport(ERROR,
1554                                 (errmsg("could not open file \"%s\": %m", path)));
1555
1556         if ((write(fd, ondisk, needed_length)) != needed_length)
1557         {
1558                 CloseTransientFile(fd);
1559                 ereport(ERROR,
1560                                 (errcode_for_file_access(),
1561                                  errmsg("could not write to file \"%s\": %m", tmppath)));
1562         }
1563
1564         /*
1565          * fsync the file before renaming so that even if we crash after this we
1566          * have either a fully valid file or nothing.
1567          *
1568          * TODO: Do the fsync() via checkpoints/restartpoints, doing it here has
1569          * some noticeable overhead since it's performed synchronously during
1570          * decoding?
1571          */
1572         if (pg_fsync(fd) != 0)
1573         {
1574                 CloseTransientFile(fd);
1575                 ereport(ERROR,
1576                                 (errcode_for_file_access(),
1577                                  errmsg("could not fsync file \"%s\": %m", tmppath)));
1578         }
1579         CloseTransientFile(fd);
1580
1581         fsync_fname("pg_llog/snapshots", true);
1582
1583         /*
1584          * We may overwrite the work from some other backend, but that's ok, our
1585          * snapshot is valid as well, we'll just have done some superflous work.
1586          */
1587         if (rename(tmppath, path) != 0)
1588         {
1589                 ereport(ERROR,
1590                                 (errcode_for_file_access(),
1591                                  errmsg("could not rename file \"%s\" to \"%s\": %m",
1592                                                 tmppath, path)));
1593         }
1594
1595         /* make sure we persist */
1596         fsync_fname(path, false);
1597         fsync_fname("pg_llog/snapshots", true);
1598
1599         /*
1600          * Now there's no way we can loose the dumped state anymore, remember
1601          * this as a serialization point.
1602          */
1603         builder->last_serialized_snapshot = lsn;
1604
1605 out:
1606         ReorderBufferSetRestartPoint(builder->reorder,
1607                                                                  builder->last_serialized_snapshot);
1608 }
1609
1610 /*
1611  * Restore a snapshot into 'builder' if previously one has been stored at the
1612  * location indicated by 'lsn'. Returns true if successful, false otherwise.
1613  */
1614 static bool
1615 SnapBuildRestore(SnapBuild *builder, XLogRecPtr lsn)
1616 {
1617         SnapBuildOnDisk ondisk;
1618         int                     fd;
1619         char            path[MAXPGPATH];
1620         Size            sz;
1621         int                     readBytes;
1622         pg_crc32        checksum;
1623
1624         /* no point in loading a snapshot if we're already there */
1625         if (builder->state == SNAPBUILD_CONSISTENT)
1626                 return false;
1627
1628         sprintf(path, "pg_llog/snapshots/%X-%X.snap",
1629                         (uint32) (lsn >> 32), (uint32) lsn);
1630
1631         fd = OpenTransientFile(path, O_RDONLY | PG_BINARY, 0);
1632
1633         if (fd < 0 && errno == ENOENT)
1634                 return false;
1635         else if (fd < 0)
1636                 ereport(ERROR,
1637                                 (errcode_for_file_access(),
1638                                  errmsg("could not open file \"%s\": %m", path)));
1639
1640         /* ----
1641          * Make sure the snapshot had been stored safely to disk, that's normally
1642          * cheap.
1643          * Note that we do not need PANIC here, nobody will be able to use the
1644          * slot without fsyncing, and saving it won't suceed without an fsync()
1645          * either...
1646          * ----
1647          */
1648         fsync_fname(path, false);
1649         fsync_fname("pg_llog/snapshots", true);
1650
1651
1652         /* read statically sized portion of snapshot */
1653         readBytes = read(fd, &ondisk, SnapBuildOnDiskConstantSize);
1654         if (readBytes != SnapBuildOnDiskConstantSize)
1655         {
1656                 CloseTransientFile(fd);
1657                 ereport(ERROR,
1658                                 (errcode_for_file_access(),
1659                                  errmsg("could not read file \"%s\", read %d of %d: %m",
1660                                                 path, readBytes, (int) SnapBuildOnDiskConstantSize)));
1661         }
1662
1663         if (ondisk.magic != SNAPBUILD_MAGIC)
1664                 ereport(ERROR,
1665                                 (errmsg("snapbuild state file \"%s\" has wrong magic %u instead of %u",
1666                                                 path, ondisk.magic, SNAPBUILD_MAGIC)));
1667
1668         if (ondisk.version != SNAPBUILD_VERSION)
1669                 ereport(ERROR,
1670                                 (errmsg("snapbuild state file \"%s\" has unsupported version %u instead of %u",
1671                                                 path, ondisk.version, SNAPBUILD_VERSION)));
1672
1673         INIT_CRC32(checksum);
1674         COMP_CRC32(checksum,
1675                            ((char *) &ondisk) + SnapBuildOnDiskNotChecksummedSize,
1676                            SnapBuildOnDiskConstantSize - SnapBuildOnDiskNotChecksummedSize);
1677
1678         /* read SnapBuild */
1679         readBytes = read(fd, &ondisk.builder, sizeof(SnapBuild));
1680         if (readBytes != sizeof(SnapBuild))
1681         {
1682                 CloseTransientFile(fd);
1683                 ereport(ERROR,
1684                                 (errcode_for_file_access(),
1685                                  errmsg("could not read file \"%s\", read %d of %d: %m",
1686                                                 path, readBytes, (int) sizeof(SnapBuild))));
1687         }
1688         COMP_CRC32(checksum, &ondisk.builder, sizeof(SnapBuild));
1689
1690         /* restore running xacts information */
1691         sz = sizeof(TransactionId) * ondisk.builder.running.xcnt_space;
1692         ondisk.builder.running.xip = MemoryContextAlloc(builder->context, sz);
1693         readBytes = read(fd, ondisk.builder.running.xip, sz);
1694         if (readBytes != sz)
1695         {
1696                 CloseTransientFile(fd);
1697                 ereport(ERROR,
1698                                 (errcode_for_file_access(),
1699                                  errmsg("could not read file \"%s\", read %d of %d: %m",
1700                                                 path, readBytes, (int) sz)));
1701         }
1702         COMP_CRC32(checksum, ondisk.builder.running.xip, sz);
1703
1704         /* restore committed xacts information */
1705         sz = sizeof(TransactionId) * ondisk.builder.committed.xcnt;
1706         ondisk.builder.committed.xip = MemoryContextAlloc(builder->context, sz);
1707         readBytes = read(fd, ondisk.builder.committed.xip, sz);
1708         if (readBytes != sz)
1709         {
1710                 CloseTransientFile(fd);
1711                 ereport(ERROR,
1712                                 (errcode_for_file_access(),
1713                                  errmsg("could not read file \"%s\", read %d of %d: %m",
1714                                                 path, readBytes, (int) sz)));
1715         }
1716         COMP_CRC32(checksum, ondisk.builder.committed.xip, sz);
1717
1718         CloseTransientFile(fd);
1719
1720         /* verify checksum of what we've read */
1721         if (!EQ_CRC32(checksum, ondisk.checksum))
1722                 ereport(ERROR,
1723                                 (errcode_for_file_access(),
1724                                  errmsg("snapbuild state file %s: checksum mismatch, is %u, should be %u",
1725                                                 path, checksum, ondisk.checksum)));
1726
1727         /*
1728          * ok, we now have a sensible snapshot here, figure out if it has more
1729          * information than we have.
1730          */
1731
1732         /*
1733          * We are only interested in consistent snapshots for now, comparing
1734          * whether one imcomplete snapshot is more "advanced" seems to be
1735          * unnecessarily complex.
1736          */
1737         if (ondisk.builder.state < SNAPBUILD_CONSISTENT)
1738                 goto snapshot_not_interesting;
1739
1740         /*
1741          * Don't use a snapshot that requires an xmin that we cannot guarantee to
1742          * be available.
1743          */
1744         if (TransactionIdPrecedes(ondisk.builder.xmin, builder->initial_xmin_horizon))
1745                 goto snapshot_not_interesting;
1746
1747
1748         /* ok, we think the snapshot is sensible, copy over everything important */
1749         builder->xmin = ondisk.builder.xmin;
1750         builder->xmax = ondisk.builder.xmax;
1751         builder->state = ondisk.builder.state;
1752
1753         builder->committed.xcnt = ondisk.builder.committed.xcnt;
1754         /* We only allocated/stored xcnt, not xcnt_space xids ! */
1755         /* don't overwrite preallocated xip, if we don't have anything here */
1756         if (builder->committed.xcnt > 0)
1757         {
1758                 pfree(builder->committed.xip);
1759                 builder->committed.xcnt_space = ondisk.builder.committed.xcnt;
1760                 builder->committed.xip = ondisk.builder.committed.xip;
1761         }
1762         ondisk.builder.committed.xip = NULL;
1763
1764         builder->running.xcnt = ondisk.builder.committed.xcnt;
1765         if (builder->running.xip)
1766                 pfree(builder->running.xip);
1767         builder->running.xcnt_space = ondisk.builder.committed.xcnt_space;
1768         builder->running.xip = ondisk.builder.running.xip;
1769
1770         /* our snapshot is not interesting anymore, build a new one */
1771         if (builder->snapshot != NULL)
1772         {
1773                 SnapBuildSnapDecRefcount(builder->snapshot);
1774         }
1775         builder->snapshot = SnapBuildBuildSnapshot(builder, InvalidTransactionId);
1776         SnapBuildSnapIncRefcount(builder->snapshot);
1777
1778         ReorderBufferSetRestartPoint(builder->reorder, lsn);
1779
1780         Assert(builder->state == SNAPBUILD_CONSISTENT);
1781
1782         ereport(LOG,
1783                         (errmsg("logical decoding found consistent point at %X/%X",
1784                                         (uint32)(lsn >> 32), (uint32)lsn),
1785                          errdetail("found initial snapshot in snapbuild file")));
1786         return true;
1787
1788 snapshot_not_interesting:
1789         if (ondisk.builder.running.xip != NULL)
1790                 pfree(ondisk.builder.running.xip);
1791         if (ondisk.builder.committed.xip != NULL)
1792                 pfree(ondisk.builder.committed.xip);
1793         return false;
1794 }
1795
1796 /*
1797  * Remove all serialized snapshots that are not required anymore because no
1798  * slot can need them. This doesn't actually have to run during a checkpoint,
1799  * but it's a convenient point to schedule this.
1800  *
1801  * NB: We run this during checkpoints even if logical decoding is disabled so
1802  * we cleanup old slots at some point after it got disabled.
1803  */
1804 void
1805 CheckPointSnapBuild(void)
1806 {
1807         XLogRecPtr      cutoff;
1808         XLogRecPtr      redo;
1809         DIR                *snap_dir;
1810         struct dirent *snap_de;
1811         char            path[MAXPGPATH];
1812
1813         /*
1814          * We start of with a minimum of the last redo pointer. No new replication
1815          * slot will start before that, so that's a safe upper bound for removal.
1816          */
1817         redo = GetRedoRecPtr();
1818
1819         /* now check for the restart ptrs from existing slots */
1820         cutoff = ReplicationSlotsComputeLogicalRestartLSN();
1821
1822         /* don't start earlier than the restart lsn */
1823         if (redo < cutoff)
1824                 cutoff = redo;
1825
1826         snap_dir = AllocateDir("pg_llog/snapshots");
1827         while ((snap_de = ReadDir(snap_dir, "pg_llog/snapshots")) != NULL)
1828         {
1829                 uint32          hi;
1830                 uint32          lo;
1831                 XLogRecPtr      lsn;
1832                 struct stat     statbuf;
1833
1834                 if (strcmp(snap_de->d_name, ".") == 0 ||
1835                         strcmp(snap_de->d_name, "..") == 0)
1836                         continue;
1837
1838                 snprintf(path, MAXPGPATH, "pg_llog/snapshots/%s", snap_de->d_name);
1839
1840                 if (lstat(path, &statbuf) == 0 && !S_ISREG(statbuf.st_mode))
1841                 {
1842                         elog(DEBUG1, "only regular files expected: %s", path);
1843                         continue;
1844                 }
1845
1846                 /*
1847                  * temporary filenames from SnapBuildSerialize() include the LSN and
1848                  * everything but are postfixed by .$pid.tmp. We can just remove them
1849                  * the same as other files because there can be none that are currently
1850                  * being written that are older than cutoff.
1851                  *
1852                  * We just log a message if a file doesn't fit the pattern, it's
1853                  * probably some editors lock/state file or similar...
1854                  */
1855                 if (sscanf(snap_de->d_name, "%X-%X.snap", &hi, &lo) != 2)
1856                 {
1857                         ereport(LOG,
1858                                         (errmsg("could not parse filename \"%s\"", path)));
1859                         continue;
1860                 }
1861
1862                 lsn = ((uint64) hi) << 32 | lo;
1863
1864                 /* check whether we still need it */
1865                 if (lsn < cutoff || cutoff == InvalidXLogRecPtr)
1866                 {
1867                         elog(DEBUG1, "removing snapbuild snapshot %s", path);
1868
1869                         /*
1870                          * It's not particularly harmful, though strange, if we can't
1871                          * remove the file here. Don't prevent the checkpoint from
1872                          * completing, that'd be cure worse than the disease.
1873                          */
1874                         if (unlink(path) < 0)
1875                         {
1876                                 ereport(LOG,
1877                                                 (errcode_for_file_access(),
1878                                                  errmsg("could not unlink file \"%s\": %m",
1879                                                                 path)));
1880                                 continue;
1881                         }
1882                 }
1883         }
1884         FreeDir(snap_dir);
1885 }